A Structured LSTM-based Network for Traffic Distribution Forecasting on Long Downhill Roads

Abstract

Long downhill road is one of a major traffic risk factor. The existing researches on road risk warning have become mature, but there is no feasible solution for effective risk warning and control on long downhill roads. The core of traffic control is the prediction of traffic distribution. To this end, this paper proposes a structured LSTM-based method for predicting the distribution of vehicles on long downhill roads. The method describes the traffic distribution based on the spatial relationship of the vehicles, and connects the LSTM network units corresponding to the directly adjacent vehicles in space with each other to share the hidden states. Thereby forming a radiation connection by multiple LSTM network units, namely the structured LSTM network unit layer. At the same time, the vehicle distribution prediction work is completed by building an encoder-decoder network architecture. The experimental results show that, compared with a single LSTM network, using a structured LSTM network to predict the vehicle distribution with bidirectional interaction can achieve better results.